AI Agents in Kenya: What They Can Do, Cost and Get Wrong

An AI agent is software that uses a language model to decide its own next step, call your tools (stock sheet, M-Pesa records, invoicing system) and keep going until a task is done. A chatbot answers; an agent acts. For most Kenyan businesses in 2026, agents work well on narrow, checkable tasks with a person approving anything that moves money or files with KRA, and they still fail when left alone on open-ended work.

Key Takeaways

• Start with one repetitive task that has a clear right answer, such as matching M-Pesa payments to orders, not “an AI that runs customer service”.
• Use a plain workflow or chatbot first if the steps never change; pay for an agent only where the path really varies.
• Keep a person approving every payment, refund, credit decision and KRA eTIMS submission, at least until months of logs show the agent is right.
• Budget for model usage per conversation, messaging fees, integration work and monitoring, not just the build.
• Give the agent the fewest permissions it needs, and log every action it takes.

What is an AI agent, and how is it different from a chatbot?

The clearest definition comes from Anthropic, the company behind Claude. In its engineering guide Building effective agents, it separates two things people lump together. Workflows are “systems where LLMs and tools are orchestrated through predefined code paths”. Agents are “systems where LLMs dynamically direct their own processes and tool usage”.

OpenAI says much the same in its Agents SDK documentation: agents are “LLMs equipped with instructions and tools”, with “a built-in loop that continues until the task is complete”. Microsoft’s Copilot Studio overview describes an agent as an assistant that follows your instructions, draws on the knowledge you connect and “uses tools to take action”.

In plain terms, there are three levels:

My view: most businesses that want “an AI agent” need a well-built workflow with one or two AI steps. Anthropic’s own advice is to find “the simplest solution possible” and notes this “might mean not building agentic systems at all”. If a vendor never mentions that option, be careful. If you have not yet settled whether you need a chatbot at all, read our guide to AI chatbot development in Kenya first.

What can AI agents reliably do in 2026, and what can they not?

Read the vendors’ own documentation rather than their launch videos. The docs are more honest.

What they do well:

• Use tools you define. Every major platform now lets an agent call functions, search your documents and connect to business systems. Google describes its Gemini Enterprise Agent Platform (the renamed Vertex AI) as a way to build agents “grounded in your enterprise data”.
• Handle variation in language. An agent can read “nimetuma 4,500 kwa till, ni ya ile order ya jana” and work out that it is a payment message about yesterday’s order, where a keyword bot would stall.
• Hand off and pause. OpenAI’s SDK supports handoffs between agents, guardrails that check inputs and outputs, and “built-in mechanisms for involving humans during agent runs”. Microsoft’s Copilot Studio workflows include “human-in-the-loop controls”.

Where they still fail:

• Assuming instead of checking. Anthropic’s computer use documentation states plainly that “Claude sometimes assumes outcomes of its actions without explicitly checking their results”. An agent that assumes a payment went through is a real problem.
• Following instructions hidden in content. The same page warns that “instructions on webpages or contained in images might override your instructions”. This is prompt injection, ranked first in the OWASP Top 10 for LLM applications. A customer message or a supplier PDF can carry it.
• Compounding errors and cost. Anthropic notes that agents “trade latency and cost for better task performance” and that autonomy brings “the potential for compounding errors”. One wrong step early on gets built upon.
• Consequential actions without a person. Anthropic recommends “asking a human to confirm decisions that might result in meaningful real-world consequences”, naming financial transactions specifically.

The industry view matches. In June 2025 Gartner predicted that over 40% of agentic AI projects would be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls, as reported by Outlook Business. The same report says Gartner estimates only about 130 of the thousands of “agentic AI” vendors are real, with many relabelling chatbots and older automation. Ask any vendor to show you the tool calls and the human checkpoint, not just the chat window.

Which AI agent platforms can a Kenyan business build on?

You do not need a platform built in Kenya. You need one that connects to the systems you already use. Here is how the four big vendors stood at the time of writing:

The platform matters less than the plumbing. As WSO2 told WSO2Con Africa 2026, businesses “cannot deploy AI agents across their operations without first connecting those agents to their existing systems”, reported by Techweez. My view: that is where most of the build effort goes in Kenya too, because M-Pesa, the stock sheet and the invoicing system rarely talk to each other.

What can an AI agent do for a Kenyan business? Three honest examples

Each of these is a worked example with stated assumptions, not a client result.

M-Pesa reconciliation

Safaricom’s Daraja portal gives developers access to M-Pesa APIs, so payment notifications can flow straight into your system. Matching those payments to orders is a classic agent job, but only part of it needs AI.

Worked example: a hardware shop in Thika takes 120 till payments a day. A staff member spends 2 minutes matching each to an order: 120 × 2 = 240 minutes, or 4 hours a day.

• Assume 75% of payments match exactly on amount and phone number: 120 × 0.75 = 90 payments. Plain rules handle these. No agent needed.
• The other 30 are messy: part payments, a spouse paying from another line, a wrong reference. The agent reads the WhatsApp thread and the open orders and proposes a match. A person approves each one in about 3 minutes: 30 × 3 = 90 minutes.
• Daily time falls from 240 to 90 minutes, a saving of 150 minutes. Plan around 120 minutes saved, not 150, because the first weeks bring more exceptions than you expect.

For what the M-Pesa side of this involves, see what a Daraja API setup involves.

WhatsApp orders

A customer writes: “Nataka bags 3 za cement na delivery Ruiru kesho.” An agent can check stock, quote from your price list, ask which part of Ruiru, draft the order and send a payment request. It should not offer discounts, extend credit or promise delivery slots it cannot see in your schedule. Those stay with your team.

You need the WhatsApp Business Platform (the API), not just the phone app, for this. Meta charges by message category and the recipient’s country code, and its pricing page changes several times a year, so check it before you budget. Our WhatsApp Business API setup guide walks through the setup.

KRA eTIMS invoices

KRA allows system-to-system integration with eTIMS through an API, using OSCU (for invoicing that is always online) or VSCU (for bulk invoicing that is not always online). The integration must be developed, tested in KRA’s sandbox and certified, either by you or a KRA-verified integrator.

So the agent’s honest role is to prepare invoice data from the order and the chat, check it against your item list, and queue it. Your certified system issues the invoice, and a person approves anything unusual. An agent that “files with KRA on its own” is the wrong design. More on getting the invoicing right in how to automate KRA eTIMS invoicing without errors.

How much does an AI agent cost to run in Kenya?

There are five cost lines: the build and integration, model usage, messaging fees (if on WhatsApp), hosting, and someone checking logs every week. We quote the build per scope after a free consultation, so I will not put a number on it here. Model usage, though, you can estimate yourself.

Worked example of model cost. Anthropic’s pricing page listed Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens at the time of writing. Assume an agent conversation makes 8 model calls, each with 4,000 input tokens and 300 output tokens.

• Input: 8 × 4,000 = 32,000 tokens. 32,000 ÷ 1,000,000 × $2 = $0.064.
• Output: 8 × 300 = 2,400 tokens. 2,400 ÷ 1,000,000 × $10 = $0.024.
• Per conversation: $0.064 + $0.024 = $0.088. At 100 conversations a day for 30 days: 3,000 × $0.088 = $264 a month.

A single-call chatbot on the same model costs 4,000 ÷ 1,000,000 × $2 + 300 ÷ 1,000,000 × $10 = $0.008 + $0.003 = $0.011 per conversation, or 3,000 × $0.011 = $33 a month. That is the agent premium: roughly eight times the model cost here, because the agent loops. Cheaper models and prompt caching narrow the gap, but budget for the loop. Then add a 30% buffer for retries and long conversations: $264 × 1.3 is about $343.

What are the risks of using AI agents, and how do you control them?

1. Prompt injection. OWASP’s first mitigation is to “restrict the model’s access privileges to the minimum necessary”. An order agent needs to read stock and write draft orders. It does not need to issue refunds.
2. Wrong actions. OWASP also advises “human-in-the-loop controls for privileged operations”. Put the approval step in the system itself, not in a staff memo.
3. Kenyan data law. Section 35(1) of the Data Protection Act, 2019 gives every data subject “a right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects concerning or significantly affects the data subject”, with exceptions set out in the Act. If your agent decides who gets credit, a loan top-up or a job interview, keep a person making the final call. Our piece on which processes AI can handle helps sort the safe tasks from the sensitive ones.
4. Runaway loops. Anthropic recommends “stopping conditions (such as a maximum number of iterations)”. Cap steps and spending per task.

What we learned running an automated AI system ourselves

Our blog and news desk run on a pipeline we built: a language model drafts each article, automated checks reject invented clients and unsourced statistics, and articles publish on a schedule. What we learned running it: an AI writer will invent statistics and client stories unless checks stop it; a human-written rule set and automated checks are what make AI content publishable. The case study has the details.

The same lesson applies to an order or payments agent. The model is the easy part. The rules, the checks and the logs are the work.

How do you start with an AI agent this week?

1. Pick one task. Choose something done many times a day with a clear right answer: matching payments, logging orders, drafting invoice lines.
2. Time it. Count how many times a day it happens and how long each takes. Write the baseline down; you will need it to judge the result.
3. Collect 50 real examples. Pull real WhatsApp messages, payment records or orders, including the messy ones and Sheng or Kiswahili ones. These become your test set.
4. Map the tools. List every system the task touches and whether it has an API. If a key system has none, fix that first.
5. Decide the human checkpoint. Write down which actions the agent may take alone and which need approval. Money, credit and tax filings start on the approval list.
6. Test before customers see it. Run the agent on your 50 examples and check every answer. Anthropic recommends “extensive testing in sandboxed environments”.
7. Go live small. Start with one branch, one till or one product line, and review the logs weekly for the first month.

Mistakes to avoid

• Buying an “agent” that is a relabelled chatbot. Ask to see the tool calls and the approval step.
• Letting the agent touch money unsupervised. One wrong refund costs more than months of staff time saved.
• Skipping the baseline. Without before numbers you cannot tell whether the agent pays for itself.
• Testing only in English. Your customers write in Kiswahili, Sheng and mixed messages. Test with those.
• Building on a product being retired. Check the vendor’s deprecation notices before you commit.

Frequently asked questions

What is the difference between an AI agent and a chatbot?

A chatbot answers questions, usually from a script or knowledge base. An AI agent decides which steps to take, calls tools such as your stock system or payment records, and works through a task until it is complete. Agents handle more variety but cost more per task and need more safeguards.

Is an AI agent the same as an AI agency?

No. An AI agent is a piece of software that performs tasks. An AI agency or consultancy, like us, is a company that helps you decide whether you need one, then designs, builds and maintains it. Many searches for “AI agency” are really looking for help choosing and building the right tool.

Can AI agents work for businesses in East Africa?

Yes. An agent can connect to the tools Kenyan firms already use, including M-Pesa through Daraja and WhatsApp through Meta’s Business Platform. The harder part is local integration and testing with Kiswahili and mixed-language messages. Test with your own customers’ real messages before trusting any vendor’s language claims.

Do I need a developer to build an AI agent?

For a simple internal agent on Microsoft 365, Copilot Studio is low-code and a capable staff member can build one. For anything connected to M-Pesa, WhatsApp or eTIMS, you need developer work, because those integrations need APIs, testing and, for eTIMS, KRA certification.

Will an AI agent replace my staff?

It is more likely to change their work than remove it. In a well-designed system, the agent clears the routine cases and your people handle exceptions, approvals and customers who need a human. Any vendor promising a fully unattended agent for payments or tax should be treated with suspicion.

The bottom line

AI agents are real and useful in 2026, but they earn their cost on narrow tasks with clear answers, good tool connections and a person approving anything that matters. Start with one task, measure it, test it on real messages and keep the approval step. If the steps never change, a simpler workflow will serve you better and cost less.

If you want a second opinion on whether your task needs an agent, a workflow or nothing at all, our free AI Business Assessment is a good place to start. Or, if a chat is easier, WhatsApp us on 0711 344 702.